Extensions/Paxa TTS for Thai, English and Mandarin
ComfyUI Extension

Paxa TTS for Thai, English and Mandarin

Text to speech in Thai, English and Mandarin Chinese from Paxa Labs: 50 voices that each speak all three, with direction tags. Speech to text with SRT or WebVTT subtitles as a research preview. Uses your own Paxa API key and credits.

By paxalabs·Created 2 days ago·Updated 2 days ago· 0
paxalabs/comfyui-paxa
Nodes—
On cloudLocal install
Stars0
Updated2 days ago
Readme

Paxa for ComfyUI

Text to speech in Thai, English and Mandarin Chinese, plus speech to text with subtitles, from Paxa Labs as ComfyUI nodes.

| Node | In | Out | |---|---|---| | Paxa Text to Speech | text | AUDIO | | Paxa Speech to Text (preview) | AUDIO | transcript, SRT or WebVTT subtitles | | Paxa Save Subtitles | subtitles | an .srt or .vtt file in ComfyUI's output folder |

  • 50 voices on Paxa TTS Flash v2, and every one speaks Thai, English and Mandarin Chinese, including sentences that mix them. Hear them at paxalabs.com/voices.
  • Direction tags in the text set the delivery, for example [whispering] or [warm, unhurried]. The tag guide is at paxalabs.com/docs/directions.
  • Subtitles that break at word boundaries, Thai included, so a cue never splits a Thai word.

The nodes call the Paxa API with your own Paxa API key and spend credits from your Paxa account. New accounts include free credits.

Install

  • ComfyUI Manager: search for "Paxa", install, and restart ComfyUI.
  • Comfy CLI: comfy node install comfyui-paxa
  • By hand: clone this repository into ComfyUI/custom_nodes, install requirements.txt with the Python that runs ComfyUI, and restart.

The only extra dependency is requests. The nodes use ComfyUI's own torch and numpy.

Your API key

  1. Create a key at paxalabs.com/app/keys.

  2. Give it to ComfyUI in one of two ways:

    • Environment variable: set PAXA_API_KEY before you start ComfyUI.

      export PAXA_API_KEY=pxa_...        # macOS and Linux
$env:PAXA_API_KEY = "pxa_..."      # Windows PowerShell

    • Key file: save the key, alone, in a file named paxa_api_key.txt in this node pack's folder, ComfyUI/custom_nodes/comfyui-paxa/. This suits the portable and desktop builds. Updates through ComfyUI Manager keep the file. On Windows, check that it isn't saved as paxa_api_key.txt.txt.

    The environment variable wins when both are set. The nodes read the key each time they run, so a new key file works without a restart.

Why there's no key field on the nodes: ComfyUI saves every node input in the workflow, and in the metadata of every image and audio file it saves. A key typed into a node would travel with every workflow and output you share.

Nodes

Paxa Text to Speech

| Input | What it does | |---|---| | text | Thai, English, Mandarin Chinese or any mix, up to 5,000 characters per run. | | voice | The voice. foithong (female) and massaman (male) come first. The last eight in the list speak on Paxa TTS Flash (v1) only: padthai, yoyo, somtam, larb, khaosoi, roti, donut and toast. | | model | paxa-tts-flash-v2, the current model, or paxa-tts-flash-v1, which has the regional Thai accents. | | speed | 0.5 to 1.5; 1.0 is the voice as designed. Every speed costs the same. | | language | How numbers, codes and symbols are read: auto, th, en or zh. Words always keep the language they're written in. | | voice_id | Optional. Any voice id, including one newer than this pack or your own designed voice (voice_...). It replaces voice when set. |

The output is mono speech, the 24 kHz WAV Paxa sends. ComfyUI reuses a node's last result while its inputs stay the same, so running the workflow again usually doesn't call Paxa. ComfyUI can drop that result, though, for example when you run another workflow in between or when it needs the memory, and then the node calls Paxa and spends credits again.

Paxa Speech to Text (preview)

It runs Paxa STT Lite, a research preview: Thai-first transcription that also handles English, Mandarin and speech that switches language mid sentence. Its behavior may change while it's in preview.

| Input | What it does | |---|---| | audio | One recording. Stereo is mixed down to mono. | | language | The language you expect: a hint, not a filter. auto detects it. | | subtitle_format | srt or vtt. | | convention | spoken writes numbers as they were said; written uses digits and symbols, as in 30%. | | vocabulary | Optional names and terms the speech is likely to contain, one per line, up to 50. |

The outputs are the transcript and the subtitle file's text. The node sends up to about 13 minutes per run; trim longer recordings with ComfyUI's Trim Audio Duration node.

Paxa Save Subtitles

Saves the subtitles under ComfyUI/output/, as .vtt for WebVTT and .srt otherwise. filename_prefix works like the one on Save Image.

Example workflows

In ComfyUI, open Browse Templates and find comfyui-paxa:

  • Paxa Text to Speech speaks a line in Thai, English and Mandarin and saves it as MP3.
  • Paxa Voiceover with Subtitles speaks a Thai script, saves it, and turns it into SRT subtitles.

When something goes wrong

The node's error message says what Paxa answered and how to fix it, with the request id to quote if you contact Paxa. For example:

  • No Paxa API key found: set the key as described above.
  • Paxa didn't accept the API key: check the key at paxalabs.com/app/keys.
  • The voice isn't available on this model: pick a voice the model speaks, or switch the model.
  • Not enough credits: nothing was charged; top up and run again.

When your plan's rate limit or concurrency limit is reached, or the model is briefly unavailable, the node waits and retries on its own. Every retry carries the same idempotency key, so a retry is never charged twice. If the connection drops while Paxa is still working, the node waits up to 5 minutes for that first attempt to finish, then takes its result without a second charge.

Connection and data

The nodes connect to one host, api.paxalabs.com, over HTTPS. Your text and audio go to Paxa to be processed, under Paxa's privacy policy. Each request names this pack and its version in an x-paxa-client header, so Paxa can count how much use comes through ComfyUI.

PAXA_BASE_URL points the nodes at another Paxa endpoint, such as a dedicated deployment. It must be an https:// address.

Development

uv sync
uv run pytest
uv run ruff check .
uv run ruff format --check .

The tests run against a local stand-in for the Paxa API and need no key. uv run python scripts/update_voices.py refreshes the bundled voice list from the public GET /v1/voices.

License

MIT. See LICENSE.